Optimal control in dynamic agri-food supply chains

Abstract

The supply chains for agriculture and food (agri-food) related products face several challenges due to uncertainties and dynamic behaviors related to fluctuations in demand, uncontrollable environmental factors, sensitive quality concerns, and profitability within a low-margin industry. This research develops data-driven stochastic models and methods for solving important problems in agri-food supply chains. Agri-food supply chains are well known to be dynamic and stochastic, yet most current models use simplified deterministic models. Instead, this research develops stochastic models and optimization methods that integrate ideas and techniques from machine learning, Big Data mining, and deep reinforcement learning to improve the supply chain performance and reduce food loss amidst many sources of uncertainty. Due to advances in computational capability and the availability of data in recent decades, it is now possible to create models with more details to better reflect the true supply chain dynamics and complexity. This research first introduces a new generalized stochastic model for representing the dynamics of complex agri-food supply chains to optimize profitability by ensuring the quality of the end-product. A specific focus is placed on the tracking of the obtained quality level throughout the steps of the supply chain since this property highly predicts if the materials in the current steps are available for use in different potential final products or the final products’ acceptability by the consumer. It is recognized that these models must be able to be developed from existing data to capture the supply chain’s complexity and quantify uncertain outcomes. This research accomplishes this by integrating data mining techniques with these models to determine the supply chain dynamics. Since deriving these dynamic behaviors from historical data can become computationally challenging, a novel approach that leverages Big Data mining tools and techniques is introduced and utilized to speed up running times without compromising the complexity or requiring more assumptions for the models. Lastly, this research analyzes how traditional techniques perform versus approximation methods for agri-food supply chain models with rolling horizons and product degradation. This can be demonstrated through a mix and blend problem, a common operation in agri-food supply chains. A new neural network architecture called OR-Net is introduced as an efficient mechanism for modeling and solving sequential integer programs such as the mix and blend problem using deep reinforcement learning. OR-Net is designed specifically to focus on the orthogonal relationships that exist between an integer program’s coefficients. Using numerical experiments, analysis is performed to evaluate the performance of OR-Net against stage-wise optimization and other approximation methods.